Remij 4,839 Posted February 28, 2020 Share Posted February 28, 2020 (edited) Those are engineering sample Data Center cards obviously... as evidenced by the low clocks and RAM.... but... it's the number of CUDA cores that's important here... because we know they will have less RAM but we also know the clocks will be much higher... likely 1.8-2.0Ghz. For reference: Turing - 4608 Cuda cores (72 CUs) @ 1.8GHz core clock: 132804 (Open CL) Ampere - 6912 Cuda cores (108 CUs) @ 1.0GHz core clock: 141654 (Open CL) Ampere - 7552 Cuda cores (118 CUs) @ 1.1GHz scores: 184096 (Open CL) Guys... Ampere will have clocks of at least 1.8Ghz... It's important to note that this is specifically compute performance.. but these are low clocked engineering samples... Already demolishing Turing running at about half the clock speed of what Ampere will run at on the consumer side... Who knows what kind of architectural improvements Ampere will have over Turing for traditional gaming workloads? ... Ampere could be a 60-75% improvement, or possibly even double the performance in certain workloads... holy fucking shit. Edited February 28, 2020 by Remij_ Link to post Share on other sites
Ike★ 2,933 Posted February 28, 2020 Share Posted February 28, 2020 MSRP $1200 Link to post Share on other sites
-GD-X★ 7,986 Posted February 28, 2020 Share Posted February 28, 2020 i have no clue what any of this means...but it sounds exciting! 1 Link to post Share on other sites
Remij 4,839 Posted February 28, 2020 Author Share Posted February 28, 2020 (edited) 15 minutes ago, Ike said: MSRP $1200 For the 3080Ti? Possibly. But I think they'll hit $999 3080 would be $799 If this holds up... the 3070 will beat a 2080ti... a mid range GPU Edited February 28, 2020 by Remij_ Link to post Share on other sites
Remij 4,839 Posted February 29, 2020 Author Share Posted February 29, 2020 ...I forgot to mention neither of those is the FULL Ampere GPU either... that would 8192 Cuda cores (128CUs) Link to post Share on other sites
Sergio Perez SP11 174 Posted February 29, 2020 Share Posted February 29, 2020 Jesus Even at a low 1650 MHz core clock that’s 25 Teraflops Link to post Share on other sites
Remij 4,839 Posted February 29, 2020 Author Share Posted February 29, 2020 1 hour ago, Carlos Vela said: Jesus Even at a low 1650 MHz core clock that’s 25 Teraflops lmao JonbHBR I told him Nvidia would have GPUs 20TF+ by the launch of the consoles. Doesn't change the fact that the Xbox is still a beast, but yea. Link to post Share on other sites
JonDnD 2,820 Posted February 29, 2020 Share Posted February 29, 2020 Yeah. We will see . Doesnt matter anyway as consoles will remain the baseline, as always Link to post Share on other sites
kaz 2,447 Posted February 29, 2020 Share Posted February 29, 2020 I have no clue what any of this means...but it sounds lame! Link to post Share on other sites
lynux3 2,250 Posted February 29, 2020 Share Posted February 29, 2020 Link to post Share on other sites
Remij 4,839 Posted February 29, 2020 Author Share Posted February 29, 2020 1 hour ago, lynux3 said: Link to post Share on other sites
jehurey 3,313 Posted February 29, 2020 Share Posted February 29, 2020 18 hours ago, Remij_ said: ...I forgot to mention neither of those is the FULL Ampere GPU either... that would 8192 Cuda cores (128CUs) Now how could they run this test with silicon that does not have all Cuda cores? Are they simply not activating all the Cuda cores in these tests? They wouldn't create a custom GPU die that has less Cuda cores than the final product, that'd be pretty pointless. Link to post Share on other sites
Remij 4,839 Posted February 29, 2020 Author Share Posted February 29, 2020 (edited) 2 hours ago, jehurey said: Now how could they run this test with silicon that does not have all Cuda cores? Are they simply not activating all the Cuda cores in these tests? They wouldn't create a custom GPU die that has less Cuda cores than the final product, that'd be pretty pointless. When you buy a GPU, you're either getting a cut down version of a die with SMs/CUs deactivated, meaning less cores, or the full die. Nvidia has different codenames for each of their dies, which serve multiple SKUs depending on the amount of SMs deactivated for yield and market purposes. TU106 (2060 and 2070), TU104 (2080), and TU102 (2080TI, Titan RTX). Their Quadro Lineup follows the same idea depending on yields. These GPUs in this test would be engineering samples of their data center series, which is why they have low clock frequencies and high RAM. These samples boards can't be utilizing the full die. Based on the 7552 cores, we can rule out that they are using 128 Cuda cores per SM and are indeed using 64. 7552 / 128 = 59 SMs... lol no.. 7552 / 64 = 118. Yes. Turing Full die (TU102): -6 GPC (Graphics Processing Clusters) -Each GPC has 12 SMs -6 x 12 = 72 SMs -64 cuda cores per SM -72 x 64 = 4608 cuda cores (note that the 2080Ti is NOT the full die. It has 4 SMs disabled) Ampere test: 7552 cuda cores 64 cuda cores per SM (CU) = 118 SMs (CUs) If you figure out how many clusters that would be... you'd get this: 118 SMs / 12 = 9.83 GPCs Which doesn't work. 118 SMs can't be the full die because it doesn't divide evenly into any realistic grouping of clusters. Which points to 128 SMs (CUs) being likely for the full chip. Full Ampere: 16 SMs per Cluster 128 SMs / 16 = 8 GPCs 128 x 64 = 8192 cuda cores Edited February 29, 2020 by Remij_ Link to post Share on other sites
Playstation Tablet 1,775 Posted March 1, 2020 Share Posted March 1, 2020 You can already feel Remij laughing at the console peasants Link to post Share on other sites
Remij 4,839 Posted March 1, 2020 Author Share Posted March 1, 2020 9 minutes ago, Team 2019 said: You can already feel Remij laughing at the console peasants Nah, not really. There's 2 reasons. First of all... I knew this was coming and said it long ago. Secondly... the consoles are going to be no slouches. Nobody is going to be calling the graphics they will be capable of as "peasant-like"... We're all gonna have some beastly systems in the future. I said before I don't think people are ready for just how good things are going to start looking very soon. It's not just about the raw power, but about the ways in which new technologies like DLSS (ML on AMD), Variable Rate Shading, Mesh Shaders, Texture-Space Shading, and consoles coming with VRR (variable refresh rate) support as standard, will be to push performance THAT much ahead. Xbox Series X is 12TF... but it's going to be even more powerful than just 2x an Xbox One X. A lot more powerful... because it's going to be far more efficient, and have many new technologies which make it perform closer to 3x, I'd say. Possibly far more depending on the workloads. They're going to be better balanced and more well rounded. I do think it will be crazy what 20TF+ Nvidia and AMD gpus will be able to do though.. especially considering that they too will have new and better technologies to be even more efficient. I just hope all things work out and these consoles and GPUs can actually release this year. Hopefully nothing gets delayed. Link to post Share on other sites
Remij 4,839 Posted March 6, 2020 Author Share Posted March 6, 2020 (edited) ... up to 2.2 FUCKING GHz Just so we're clear... that's... 36TF Holy FUCK... I told yall motherfuckers Nvidia ain't playing around. Obviously the 3080ti is going to be cut down from that... but we're looking at 25TF+ I'd bet. @DynamiteCop! Dark Edited March 6, 2020 by Remij_ Link to post Share on other sites
Alphonse★ 454 Posted March 7, 2020 Share Posted March 7, 2020 damn, can't wait for that 3080 ti. I's about time I put my 980 ti to rest Link to post Share on other sites
Ike★ 2,933 Posted March 7, 2020 Share Posted March 7, 2020 I hope they have a pretty tech demo to show the power of the card. 1 Link to post Share on other sites
Sergio Perez SP11 174 Posted March 7, 2020 Share Posted March 7, 2020 That’s 36 Teraflops what in tarnation Link to post Share on other sites
Recommended Posts
Please sign in to comment
You will be able to leave a comment after signing in
Sign In Now